White House Restricts AI Model Sharing with UK Safety Institute
The White House has instructed OpenAI and Anthropic to withhold new AI models from the UK's AI Safety Institute until US agencies complete their own reviews, raising questions about international cooperation in AI safety testing.
The White House has directed leading artificial intelligence companies OpenAI and Anthropic to refrain from sharing their newest AI models with the United Kingdom's AI Safety Institute until United States agencies have had the opportunity to review them first. The instruction, reported by The Decoder, marks a significant shift in the dynamics of international AI safety collaboration and underscores growing tensions over who gets early access to cutting-edge AI technology.
According to the report, the White House wants American agencies to conduct their own evaluations of new models before those models are made available to British testers. This effectively places a US-first review process ahead of the existing arrangement with the UK's AI Safety Institute, which was established to independently test and assess the safety of advanced AI systems. The directive applies specifically to OpenAI and Anthropic, two of the most prominent US-based AI developers.
The UK's AI Safety Institute, launched in November 2023, was the first government-backed body dedicated to evaluating the safety of frontier AI models. It has worked with major AI companies to conduct pre-deployment testing and has been viewed as a model for international cooperation on AI safety. The new US instruction could complicate that collaboration by introducing a layer of American oversight before British experts can access the models.
The White House's move reflects a broader concern within the US government about maintaining technological leadership and ensuring that advanced AI systems are thoroughly vetted by American authorities. It also signals a desire to control the flow of sensitive AI technology, even to close allies. The directive does not prohibit sharing entirely, but it establishes a sequence: US review first, then international sharing.
For OpenAI and Anthropic, the instruction creates a delicate balancing act. Both companies have publicly committed to safe AI development and have worked with international bodies, including the UK's AI Safety Institute. They now must navigate between complying with US government expectations and maintaining their global partnerships. Neither company has issued a public statement on the matter.
The development comes amid a wider global conversation about AI governance. The UK has positioned itself as a leader in AI safety, hosting the first global AI Safety Summit in Bletchley Park in November 2023. The US, meanwhile, has pursued its own regulatory path, including an executive order on AI issued in October 2023 that required developers to share safety test results with the government. The new directive appears to extend that logic to international sharing.
Critics may view the White House's instruction as a setback for international cooperation on AI safety, which many experts argue is essential given the borderless nature of AI risks. Supporters, however, may see it as a necessary step to ensure that US agencies have full visibility into frontier models before they are deployed or shared abroad. The long-term impact on the UK's AI Safety Institute and its ability to attract top talent and conduct meaningful evaluations remains uncertain.
The report does not specify which US agencies will conduct the reviews or what criteria they will use. It also does not indicate whether other AI companies, such as Google DeepMind or Meta, have received similar instructions. The White House has not publicly commented on the directive.
As AI development continues to accelerate, the episode highlights the tension between national security interests and the global coordination needed to manage AI risks. For now, the UK's AI Safety Institute will have to wait for US agencies to complete their reviews before gaining access to the latest models from OpenAI and Anthropic.
6
